This is a repository containing code for a parallelized Sobel algorithm that utilizes MPI and CUDA function calls. This program is meant to run on Power9 system architecture.
PARALLEL SOBEL BOUNDARY DETECTION
This is a program that runs the Sobel boundary detection algorithm using MPI, MPI_IO, and CUDA functions.
Sobel.c contains all MPI code and the main function.
Sobel.cu contains Cuda implementation.
pixel.h contains 2 datatypes that are used in both Sobel.c and Sobel.cu
There is a Makefile included that will compile all 3 of these files into a simple binary.
There are also 4 sbatch scripts that cumulatively test strong and weak scaling using the BlackMarble 2016 photograph.
The photograph used for testing can be found here: https://svs.gsfc.nasa.gov/30876
The files used for testing are much too large for upload to GitHub, but output images will be present within the write-up.
To run the binary, the parameters look something like:
mpirun -np RANKS ./Sobel-mpi-multi-gpu WIDTH HEIGHT input-filename output-filename
Be wary that photographs must be in a .data (byte) format.
Use an RGB file where each color value is a byte and ordered contiguously. GIMP is recommended for this, but not entirely necessary.
The output-filename is whatever you choose to name your output file.
This output file will be a grayscale representation of the calculated Sobel valued pixels. Thus, it will outline any boundaries found within the input photograph.
To view your output file GIMP is again recommended. It will ask for various specifications:
- Make sure your height and width is set to the same height and width of your input file
- Offset must be 0
- Image Type should be Gray 32-bit Float
It is also recommended to go into GIMP color>Auto>Stretch Contrast for better viewing
These are all testing scenarios present within the sbatch scripts.
Strong Scaling:
BlackMarble_2016_C1_18432x18432 1 ranks 1 node
Total Runtime: 8.726945 seconds
Cuda Runtime: 5.348938 seconds
BlackMarble_2016_C1_18432x18432 2 ranks 1 node
Total Runtime: 5.795312 seconds
Cuda Runtime: 2.884574 seconds
BlackMarble_2016_C1_18432x18432 4 ranks 1 node
Total Runtime: 3.791389 seconds
Cuda Runtime: 1.432372 seconds
BlackMarble_2016_C1_18432x18432 8 ranks 2 node
Total Runtime: 4.009085 seconds
Cuda Runtime: 0.745842 seconds
BlackMarble_2016_C1_18432x18432 16 ranks 3 node
Total Runtime: 4.296658 seconds
Cuda Runtime: 0.469719 seconds
BlackMarble_2016_C1_18432x18432 32 ranks 6 node
Total Runtime: 3.910650 seconds
Cuda Runtime: 0.285459 seconds
Weak Scaling:
BlackMarble_2016_Full_11585x11585 1 ranks 1 node
Total Runtime: 3.669618 seconds
Cuda Runtime: 2.245136 second
BlackMarble_2016_Full_16384x16384 2 ranks 1 node
Total Runtime: 4.811697 seconds
Cuda Runtime: 2.360394 seconds
BlackMarble_2016_Full_23170x23170 4 ranks 1 node
Total Runtime: 7.116398 seconds
Cuda Runtime: 1.920250 seconds
BlackMarble_2016_Full_32768x32768 8 ranks 2 nodes
Total Runtime: 11.498551 seconds
Cuda Runtime: 1.904782 seconds
BlackMarble_2016_Full_46341x46341 16 ranks 3 nodes
Total Runtime: 23.138975 seconds
Cuda Runtime: 2.169928